A facial-login system unlocks reliably for light-skinned employees but rejects many dark-skinned employees. Its training photos came mostly from one office whose staff were light-skinned. What data property best explains the gap?

A school uses a model to flag essays for possible plagiarism. It flags English learners far more often after being trained mostly on essays by native English speakers. Which change would have best prevented this bias?

A hospital model estimates who is likely to miss an appointment. It is not given income, but it is given each patient's ZIP code; in the area, ZIP codes closely track neighborhood income. What is the ZIP code acting as?

A retailer trains a fraud model only on purchases its team had investigated. That team had investigated one neighborhood far more often than others. The model now flags that neighborhood's purchases most often. What in the training data explains this?

A speech-to-text app was built from recordings of adults in quiet studios. It works poorly for children speaking in busy classrooms. Which addition to the training data would best have prevented the problem?

A health-risk classifier gives lower-risk predictions to a group even when their symptoms and test results are similar to other patients. Its data came from clinics that gave that group fewer diagnostic tests. What kind of bias is in this data set?

A company trains a resume-screening model on employees it promoted over 15 years, having historically promoted very few women into technical roles. The model rates resumes like past male hires' more highly. What in the data explains this?

A model predicts which students should receive tutoring. It was trained only on students who had already asked for help, not on all students who might need it. What data issue is most likely?

A navigation app predicts travel times well in wealthy areas but poorly in rural areas. Its data came from drivers who used the company's newest phones, which were uncommon in rural areas. What most directly caused the difference?

An AI tool recommends advanced courses based on past enrollments. Because some schools had not offered advanced courses, their students rarely appear among past enrollees. What change would most directly prevent the recommendation bias?

A photo-tagging model labels images of doctors correctly much more often when the doctor is a man. Its training set has many more photos of male doctors than female doctors. Which term best names this problem?

A city uses a model trained on past reports to decide where to inspect housing. Residents in one area had less access to the reporting website, so few problems from that area appear in the records. What consequence is most likely after deployment?

A hiring model was trained on successful employees from one company, where nearly all managers had attended a small group of colleges. It rejects applicants from other colleges with similar job skills. Which data change would best reduce this bias?

A loan model excludes race but uses neighborhood, school district, and property values, which are strongly linked to racial segregation. The team removes the race column but changes nothing else. Why does the bias remain?

A voice assistant is trained with recordings from speakers in four regions, but almost all recordings have the same accent within each region. It frequently misunderstands speakers with other accents from those regions. What would have prevented this?

A college-admissions model learns from past decisions. Applicants who needed fee waivers were historically admitted at lower rates despite similar records, and the model keeps giving them lower recommendations. What type of bias does this show?

A team trains a model to recognize skin conditions using photos taken mostly under bright clinic lighting. It performs poorly on photos taken in homes with dim lighting. What training-data property explains this result?

An online store uses a model trained on past purchases to decide which customers see discount offers. Customers without reliable internet bought less often online, so they receive fewer offers. What is the likely consequence?

A school wants an AI model to predict which students may need transportation support. The data include students who rode the bus last year but omit students who could not reach a bus stop. Which choice best prevents bias before training?

A product-safety model is trained on customer complaints, mostly in English since the reporting form is English-only. It finds fewer problems in products used by non-English speakers. What kind of bias does this show?

A medical triage model trained on hospital records ranks patients without transportation as lower priority, since their records show fewer past visits. What data source should be added before deployment?

A district removes names and gender fields from a discipline-referral model but keeps teacher comments and past referrals, which reflect different treatment of students by gender. Which step would best reduce the bias before retraining?

A translation system is trained mostly on formal documents written by adults. It produces poor translations for messages written by teenagers using everyday slang. What data problem is shown?

A predictive-policing model is trained on arrest records from neighborhoods that were patrolled most heavily. It recommends sending more officers to those same neighborhoods. What feedback effect can result?

A benefits-screening model favors applicants with long credit histories. Designers remove that field but keep bank balance, homeownership, and years at address. What is the best way to check whether removing the field fixed the problem?